MLtraq

What it does

Artifact storage

The FAQ says arbitrary objects, including model weights and state parameters, can be dumped and loaded, with filesystem storage or third-party services and data stores.checked 8 Oct 2026

Collaboration

Users can back up, merge, share, and reload experiments with their computation state.checked 8 Oct 2026

Compatibility caveat

The documentation says interfaces may change as MLtraq progresses and recommends pinning an exact version and verifying updates with tests.checked 8 Oct 2026

Database integrations

The default database is SQLite, and MLtraq can connect to SQL databases supported by SQLAlchemy.checked 8 Oct 2026

Distribution

The installation instructions use pip install mltraq --upgrade.checked 8 Oct 2026

Execution

It supports experiment runs with parameter grids, chained steps, parallel execution, and resuming from persisted state.checked 8 Oct 2026

Experiment state

It can track experiment state, stream metrics, reproduce runs, collaborate, and resume computation.checked 8 Oct 2026

Interfaces

Experiments can be accessed with Python, Pandas, and SQL from Python scripts, Jupyter notebooks, and dashboards.checked 8 Oct 2026

Interoperability

Experiments can be accessed with Python, Pandas, and SQL using native database types and open formats.checked 8 Oct 2026

License

The project is licensed under the BSD 3-Clause License.checked 8 Oct 2026

Metrics

Experiments can stream metrics and track native Python data types and structures, as well as NumPy, Pandas, and PyArrow objects.checked 8 Oct 2026

Model support

MLtraq is model-agnostic, and its FAQ describes adding a PyTorch model through a scikit-learn-compatible wrapper or a redesigned training step.checked 8 Oct 2026

Parallel backends

The documentation names Dask, Ray, Spark, and custom cluster-specific backends as options for computation.checked 8 Oct 2026

Parallel execution

Chained execution uses joblib.Parallel with process-based parallelism, and Dask, Ray, Spark, or custom backends can be used.checked 8 Oct 2026

Purpose

MLtraq is an open-source Python library for ML and AI developers to design, execute, track, and share experiments.checked 8 Oct 2026

Requirements

The documented requirements are Python 3.9+, SQLAlchemy 2.0+, Pandas 1.5.3+, and Joblib 1.3.2+.checked 8 Oct 2026

Security

MLtraq's DATAPAK format uses a subset of safe Python Pickle opcodes for listed basic and container types.checked 8 Oct 2026

Security detail

The state-storage documentation says its DATAPAK serialization uses only a subset of safe Pickle opcodes for listed basic and container types.checked 8 Oct 2026

Storage

The DataStore interface stores objects outside the database and uses the filesystem by default; third-party storage options can be requested through the issue tracker or discussion board.checked 8 Oct 2026

Support

The project directs users to GitHub Discussions to ask questions and share ideas.checked 8 Oct 2026

Supported data

It supports native Python types and structures, as well as NumPy, Pandas, and PyArrow objects.checked 8 Oct 2026

Target users

The FAQ describes MLtraq as a good candidate for experimentation and says MLflow may be a better fit for users prioritizing MLOps.checked 8 Oct 2026